Why AI Projects Fail, and How to Succeed

You've seen the number. "95% of AI projects fail." It's in the keynotes, the LinkedIn threads, the "AI is just a bubble" takes - repeated far more often than it's actually read (opens in new tab). And if you run a business and you've been cautiously circling AI - maybe your team has ChatGPT, maybe you tried a tool and nothing really changed - that number is the perfect excuse to wait. If nineteen out of twenty projects flop, why risk yours?

Here's the problem with that conclusion. We went back to where the stat actually comes from. The number is real, but it doesn't say what the headline says. Read properly, it's not a reason to wait. It's a map of exactly how to not waste your money.

Where the number actually comes from

It traces to a single source: MIT Media Lab's Project NANDA report, The GenAI Divide: State of AI in Business 2025. (opens in new tab) Almost everyone repeating "95%" is quoting a headline about it, not the report itself.

And what it measured matters. The study judged AI pilot outcomes by a single yardstick - measurable impact on profit and loss. It didn't find that 95% of AI models broke or didn't work. It found that about 5% of pilots drove rapid revenue acceleration, (opens in new tab) while the vast majority delivered little to no measurable impact on profit and loss. That's a measure of AI value realization, not technology failure. The AI usually ran fine. It just didn't move the business - which is a very different, and far more fixable, problem.

The number is messier than the headline

Once you look closely, the clean "95%" gets blurry.

It's drawn from a subset. The failure figure comes specifically from custom, task-specific GenAI tools - meanwhile around 40% of companies that piloted general-purpose LLMs successfully got them into production (opens in new tab). Even the study's own methodology gets reported inconsistently across outlets - some cite 150 executive interviews, (opens in new tab) 350 employee surveys, and 300 deployments, others cite far smaller numbers. And critics have pushed on the rigor: one widely-shared critique argued that if MIT stands behind the claims (opens in new tab) it should release the full supporting data, and if not, retract the report.

None of this means the study is worthless. It means the honest version of the stat is "most enterprise GenAI pilots haven't yet produced measurable financial returns" - which is true, sobering, and a lot less clickable than "95% fail."

Consider who's selling the cure

There's one more thing worth knowing. The report concludes that the fix is more agentic AI - built on protocols that NANDA itself develops. Even Fortune (opens in new tab), which covered the study seriously, added the caveat: NANDA has an incentive to suggest current methods aren't working, so it's worth considering the source.

That doesn't make the findings wrong. It's just a reminder that the loudest AI statistics often come from someone with something to sell - including, fairly, vendors like us. Which is exactly why the next part matters more than the number.

The real reason AI projects fail

Skip the headline and read what the study says about why pilots stalled. It's the most valuable thing in the report, and it gets repeated far less than the number itself.

When AI projects fail, it's rarely the model that broke. MIT described a "learning gap" - organizations didn't know how to design workflows that capture AI's value. In other words, the tool worked; the process around it didn't exist. That lines up with what every serious analysis of these failures keeps finding: the failure is rooted not in flawed models but in poor integration and misaligned priorities (opens in new tab).

Then there's the misallocation, and this one is worth reading twice: more than half of AI budgets go to sales and marketing tools, yet MIT found the biggest ROI in back-office automation (opens in new tab) - eliminating outsourcing, cutting agency costs, and streamlining operations. Companies chase the flashy customer-facing use case and underfund the boring internal one that actually pays.

And finally, on how the winners got there: buying from specialized vendors and building partnerships succeeded about 67% of the time, (opens in new tab) while internal builds succeeded only one-third as often.

What this actually means if you're on the fence

Put those three findings together and the picture flips. Most AI projects fail not because the technology is weak, but because of how they're chosen and run. AI project success didn't come down to better models - everyone has access to the same ones. The 5% who got there did three unglamorous things: they picked a process worth automating, they defined what a win looked like before building, and they integrated the tool into how the team already works instead of bolting it on the side.

The 95% bought a tool and hoped.

We know this pattern because we hit it on ourselves. Our own back-office was the mess - invoicing split across five tools that didn't talk to each other. We didn't start with "what AI can we add." We mapped where the time actually went, then built one system around it (opens in new tab): 40% faster processing, 30% lower spend on the tools it replaced, and 10x the volume with no extra headcount. Boring, internal, back-office - and exactly the category MIT says delivers the highest return. We only took it to clients once it worked for us.

How to land in the 5%

If there's one practical takeaway from a much-abused statistic, it's this: don't start with AI. Start with the work.

Map where your team's hours actually go. Find the one process that quietly eats time every week and runs mostly on manual judgment you could codify - usually it's invoicing, reporting, or data being copied between systems. Define what success looks like in plain numbers before anyone builds anything. Then automate that one thing, integrated into your existing stack, and prove the number before you scale. AI projects fail when teams skip that and start with the tool instead.

That's the whole difference between the 5% and the 95%. Not the model. The process underneath it.

That mapping step is exactly what an AI audit (opens in new tab) is for - an honest look at where automation would actually pay off, and where it wouldn't. Sometimes the answer is a system. Sometimes it's a simpler fix, and sometimes it's "you don't need AI here yet." We'd rather tell you that than sell you into the 95%.